What is Time-Varying Confounding?
Time-Varying Confounding is confounding in longitudinal data where a variable predicts later treatment and outcome, may change across follow-up, and may itself be affected by earlier treatment.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Exposure-confounder feedback creates the hard case
Suppose prior therapy improves a biomarker, the biomarker guides the next dose, and both influence the final outcome. The current biomarker must be handled to compare later doses, but ordinary adjustment changes the pathway through which earlier therapy acts. A time-varying covariate is not automatically a confounder: its temporal role must be justified with the treatment, covariate, and outcome timeline.
Longitudinal G-methods target explicit strategies
The parametric G-formula simulates outcomes under specified treatment rules, Marginal Structural Models commonly use sequential inverse-probability weights, and structural nested models use G-estimation. Each approach needs Consistency, sequential Exchangeability, Positivity, correct temporal ordering, and valid treatment and outcome definitions. Daniel and colleagues explain why conventional adjustment can be biased and compare the main G-methods.
Diagnostics must follow every treatment decision
Inspect conditional treatment probabilities at each time, stabilized-weight distributions, effective sample size, censoring weights, treatment-history counts, and support for the strategies being compared. Extreme weights often reveal practical Positivity problems or model error. Truncation can reduce variance but changes the estimator and does not create missing treatment histories, so report it with sensitivity analyses.
Key Characteristics
- Requires a longitudinal treatment-covariate-outcome timeline
- Includes covariates that confound later treatment decisions
- Can involve covariates affected by earlier treatment
- Makes baseline-only adjustment insufficient for sequential strategies
- Motivates G-methods rather than ordinary time-updated regression
- Requires time-specific overlap, model, censoring, and weight diagnostics
Common Use Cases
- Estimating sustained medication strategies with changing disease severity
- Evaluating adaptive notification policies using evolving engagement signals
- Comparing treatment sequences when earlier actions alter later risk scores
- Designing longitudinal Target Trial Emulations with treatment switching
- Auditing whether a post-treatment feature is a mediator, confounder, or collider
Example
Loading code...Frequently Asked Questions
How is Time-Varying Confounding different from ordinary confounding?
Ordinary point-treatment confounding can often be represented by baseline common causes of treatment and outcome. Time-varying confounding involves repeated decisions and evolving covariates; the covariate before a later decision may also have been changed by an earlier treatment, creating exposure-confounder feedback.
Why not put every time-varying covariate into a regression model?
A covariate affected by prior treatment can mediate part of that treatment's effect while confounding a later decision. Conditioning on it in an ordinary outcome regression can remove part of the effect or induce selection bias. The correct strategy depends on the longitudinal estimand and causal structure.
Which methods handle Time-Varying Confounding?
Common choices are the parametric G-formula, Marginal Structural Models estimated with inverse-probability weights, and structural nested models estimated with G-estimation. They encode different models and failure modes, but all require a coherent intervention, sequential assumptions, and adequate support.
What does sequential exchangeability mean here?
At each treatment decision, the action must be independent of future counterfactual outcomes after conditioning on the measured history available before that decision. This is stronger than baseline exchangeability and cannot be verified from observed data alone; omitted time-varying causes can invalidate it.
How should a Time-Varying Confounding analysis be diagnosed?
Draw the temporal causal structure, verify timestamps, inspect treatment probabilities and history counts at every decision, summarize stabilized and censoring weights, calculate effective sample size, test alternative nuisance models and truncation thresholds, and perform sensitivity analysis for unmeasured confounding.